An open-source multi-step deep research agent that decomposes questions, recursively crawls primary web sources in parallel, cross-verifies facts, and synthesizes publication-grade intelligence dossiers with in-text citations.
💡
Universal Model Support
Deep Dossier works seamlessly with Google Gemini, Anthropic Claude, OpenAI GPT, DeepSeek, NVIDIA Nemotron, and local models via Ollama or vLLM.
Overview
While proprietary systems lock autonomous deep research behind $200/month subscriptions, Deep Dossier provides an open-source, local terminal agent. It combines targeted hypothesis planning, zero-config web crawling, and large-context synthesis to deliver publication-ready documentation in minutes.
Quick Start
Install Deep Dossier directly from PyPI:
Terminal — Install from PyPI
pip install deep-dossier
Authentication
Export any available API key in your shell:
Terminal — Environment Setup
# Google Gemini (Default)
export GEMINI_API_KEY="AIzaSy..."
# Or Anthropic Claude
export ANTHROPIC_API_KEY="sk-ant-..."
# Or OpenAI
export OPENAI_API_KEY="sk-proj-..."
# Or DeepSeek
export DEEPSEEK_API_KEY="sk-..."
Running a Research Mission
Terminal — Execution
# Basic research prompt
deep-dossier "Analyze battery degradation patterns of LFP vs NMC cells in commercial EVs"
# Deep dive with Claude 3.5 Sonnet
deep-dossier "Quantum error correction thresholds" --provider anthropic --model claude-3-5-sonnet-20241022 --depth 3
# Local offline execution with Ollama
deep-dossier "Edge AI deployment architectures" --base-url http://localhost:11434/v1 --model llama3.2
GitHub Actions CI/CD Integration — Use in Any Project
Deploy automated deep research directly into any repository's CI/CD pipeline. Deep Dossier is published as a reusable composite action on the GitHub Marketplace:
⚡
Marketplace Action Identifier
Reference in your workflows via uses: shubhransh-gupta/deep-dossier@v1.
Step 1: Configure AI Provider Secret
In your target GitHub repository, go to Settings → Secrets and variables → Actions → New repository secret.
Add your API key secret (e.g. GEMINI_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY, or DEEPSEEK_API_KEY).
Step 2: Add Workflow to Your Repository
Create .github/workflows/deep-research.yml in your project:
YAML — .github/workflows/deep-research.yml
name: Automated Deep Research
on:
workflow_dispatch:
inputs:
topic:
description: 'Topic or investigative question to research'
required: true
default: 'Frontier AI reasoning architectures and benchmark trends'
depth:
description: 'Search depth (1-3)'
required: false
default: '2'
schedule:
# Optional: Run every Monday morning at 09:00 UTC
- cron: '0 9 * * 1'
jobs:
research:
runs-on: ubuntu-latest
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: Run Deep Dossier
uses: shubhransh-gupta/deep-dossier@v1
with:
topic: ${{ inputs.topic || 'Weekly frontier AI industry research' }}
api-key: ${{ secrets.GEMINI_API_KEY }}
provider: 'gemini'
depth: ${{ inputs.depth || '2' }}
output-dir: './research_outputs'
- name: Upload Research Dossier Artifacts
uses: actions/upload-artifact@v4
with:
name: research-dossier
path: ./research_outputs/
Action Inputs
Input
Required
Default
Description
topic
Yes
—
The research question or investigative topic to explore.
api-key
Yes
—
API key for the selected AI provider (passed via GitHub Secrets).
provider
No
gemini
AI provider: gemini, anthropic, openai, or deepseek.
model
No
Auto
Specific model string (e.g. gemini-3-flash-preview, claude-3-5-sonnet-20241022).
depth
No
2
Recursive search depth (1-3 iterations).
breadth
No
4
Number of sources to retrieve per query.
output-dir
No
./research_outputs
Directory where Markdown and HTML dossiers are written.
Command-Line Reference
Option
Type
Default
Description
topic
String
Interactive prompt
The core question or subject matter to investigate.
--depth, -d
Integer
2
Number of recursive search and hypothesis refinement loops.
--breadth, -b
Integer
4
Number of search results to retrieve and crawl per iteration.
Explicit model identifier (e.g. claude-3-5-sonnet-20241022, gpt-4o).
--base-url
String
None
Custom endpoint for Ollama (http://localhost:11434/v1) or vLLM.
--output, -o
Path
./research_outputs
Directory where Markdown, HTML, and JSON dossiers are saved.
Frequently Asked Questions (FAQ)
Proprietary services lock research capabilities behind closed cloud accounts and high monthly subscription tiers. Deep Dossier is open-source, local, and supports your own API keys. Furthermore, it allows you to swap models (Gemini, Claude, GPT, DeepSeek, or local Ollama) and inspect the raw JSON data and scraped documents directly on your machine.
No. Deep Dossier comes with an integrated, zero-configuration search client powered by DuckDuckGo and fallback HTTP resolvers. You only need your chosen LLM API key (or a local Ollama instance).
Yes! By passing --base-url http://localhost:11434/v1 --model llama3.2 (or your preferred local model in Ollama/vLLM), the agent's cognitive loops run entirely on your local GPU/CPU.
Every completed research run generates three synchronized files:
Markdown (.md): Publication-ready technical report formatted with numbered citations.
Interactive HTML (.html): Dark-mode responsive briefing document with clickable bibliography cards.
JSON Dossier (.json): Full raw text excerpts, timestamps, and metadata for programmatic workflows.
An autonomous test-driven code healing and bug patching agent that diagnoses issues, locates bug sites, iteratively modifies files, runs local test suites, and uses error stack traces to self-correct until all tests pass and changes are committed to Git.
🛡️
Safe Branch Isolation
Patch Pilot creates an isolated Git branch (e.g. fix/patch-pilot-...) for all modifications, ensuring your working branch is never overwritten unexpectedly.
Overview
Patch Pilot acts as an autonomous pair programmer in your local terminal. Unlike passive autocomplete plugins, Patch Pilot operates in an active test-driven loop: it analyzes bug reports, identifies candidate files, applies patches, runs your project's test suite, and feeds back stack traces until all tests are green.
Quick Start
Install Patch Pilot directly from PyPI:
Terminal — Install from PyPI
pip install patch-pilot-ai
Running Inside a Repository
Navigate to any project directory with existing tests and invoke Patch Pilot:
Terminal — Execution
cd my-project/
# Automatic test framework detection
patch-pilot "Fix KeyError when parsing optional auth headers in client.py"
# Targeted test suite run with Claude 3.5 Sonnet
patch-pilot "Resolve race condition in worker pool" \
--provider anthropic \
--model claude-3-5-sonnet-20241022 \
--test-cmd "pytest tests/test_pool.py"
# Local model via Ollama
patch-pilot "Fix off-by-one index error" \
--base-url http://localhost:11434/v1 \
--model qwen2.5-coder
Supported Test Frameworks
Ecosystem
Marker Files
Default Test Command
Python
pytest.ini, pyproject.toml, tests/
pytest
Python (Standard)
setup.py
python -m unittest discover
JavaScript / TypeScript
package.json
npm test
Rust
Cargo.toml
cargo test
Go
go.mod
go test ./...
GitHub Actions CI/CD Integration — Use in Any Project
Automatically diagnose failing tests and self-heal pull requests or broken builds in any repository. Patch Pilot is published as a reusable composite action on the GitHub Marketplace:
🤖
Marketplace Action Identifier
Reference in your workflows via uses: shubhransh-gupta/patch-pilot@v1.
Step 1: Configure AI Provider Secret
In your target GitHub repository, go to Settings → Secrets and variables → Actions → New repository secret.
Add your API key secret (e.g. GEMINI_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY, or DEEPSEEK_API_KEY).
Step 2: Add Workflow to Your Repository
Create .github/workflows/patch-pilot-ci.yml in your project:
YAML — .github/workflows/patch-pilot-ci.yml
name: Patch Pilot Self-Healing CI
on:
workflow_dispatch:
issue_comment:
types: [created]
jobs:
heal:
# Triggers manually or when any PR comment contains '/heal'
if: contains(github.event.comment.body, '/heal') || github.event_name == 'workflow_dispatch'
runs-on: ubuntu-latest
permissions:
contents: write
pull-requests: write
steps:
- name: Checkout Code
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Self-Heal Codebase with Patch Pilot
uses: shubhransh-gupta/patch-pilot@v1
with:
api-key: ${{ secrets.GEMINI_API_KEY }}
provider: 'gemini'
issue: 'Fix failing unit tests in test suite'
max-attempts: '4'
Action Inputs
Input
Required
Default
Description
api-key
Yes
—
API key for the selected AI provider (passed via GitHub Secrets).
provider
No
gemini
AI provider: gemini, anthropic, openai, or deepseek.
model
No
Auto
Specific model string (e.g. gemini-3-flash-preview, claude-3-5-sonnet-20241022).
issue
No
Diagnose and fix...
Issue or bug description for the agent to resolve.
test-command
No
Auto
Custom test command (e.g. pytest tests/test_auth.py or npm test).
max-attempts
No
4
Maximum iterative healing cycles before terminating.
Command-Line Reference
Option
Type
Default
Description
issue
String
Interactive prompt
Natural language description of the bug or ticket.
--repo, -r
Path
. (Current dir)
Path to the target Git repository to inspect and patch.
--test-cmd, -t
String
Auto-detected
Explicit test runner command (e.g. pytest tests/test_api.py).
--max-attempts
Integer
4
Maximum number of iterative self-healing cycles.
--provider, -p
String
Auto-detected
anthropic, openai, gemini, deepseek, ollama.
--model, -m
String
Provider default
Specific model string (e.g. claude-3-5-sonnet-20241022, gpt-4o).
--no-branch
Flag
False
Disables automatic creation of a dedicated Git fix branch.
Frequently Asked Questions (FAQ)
Patch Pilot automatically branches off your current Git HEAD into a isolated branch (e.g., fix/patch-pilot-issue-name). Furthermore, changes are only staged and committed when your project's test suite returns a clean exit code 0.
Rather than giving up, Patch Pilot extracts the test error summary and stack trace, injects them back into the model's prompt, and instructs it to analyze the exact failure and generate a revised patch. This loop continues until tests pass or the max attempts threshold is reached.
Yes. Patch Pilot auto-detects Node.js (npm test), Rust (cargo test), and Go (go test ./...). You can also pass any custom test command via --test-cmd "make test".
Anthropic Claude 3.5 Sonnet (--provider anthropic --model claude-3-5-sonnet-20241022) and OpenAI GPT-4o consistently show top tier performance on complex refactoring. For fast local workflows, qwen2.5-coder via Ollama provides strong offline code editing.
Inspired by the beloved Harry Potter magical creature with an insatiable appetite for finding shiny gold, Niffler sniffs out the absolute lowest airfares, verified boutique stays, IRCTC Tatkal train berths, Form-16 tax savings, and compounding EPF wealth.
🪙
Zero CA Fees & Safe Payment Handoff
Niffler never touches your banking or UPI credentials. It prepares your bookings and tax manifests, optimizes your liability, and lands you directly on the authenticated merchant gateway to complete payments and filings safely.
Niffler tackles everyday money-draining hassles for working professionals:
Travel Fare Radar: Scans across IndiGo, Air India, Akasa Air, and SpiceJet to unearth the cheapest flight fares, discovers verified boutique stays, tracks IRCTC train berths with Tatkal timing alerts, and auto-fills passenger manifests directly onto checkout.
Autonomous Citizen FinTech: Computes exact tax under both the New Tax Regime (Section 115BAC with Budget standard deduction of ₹75,000 and ₹7.75L zero-tax rebate) and Old Tax Regime (80C/80D/HRA), generates schema-valid ITR-1 Sahaj JSON, and tracks EPF balances with compound interest projections.
Quick Start
Install Niffler directly from GitHub or source:
Terminal — Install & Run
git clone https://github.com/shubhransh-gupta/niffler.git
cd niffler
pip install -e .
Launch Interactive Web Dashboard
Terminal — Start Web Portal
niffler web --port 8501
# Open http://localhost:8501 in your browser
Command-Line Usage Examples
Terminal — Travel & Tax CLI
# 1. Search cheapest flights between cities
niffler travel flights DEL BLR --direct
# 2. Discover hotels and homestays with budget and couple-friendly filters
niffler travel hotels GOA --budget 6000 --couple
# 3. Check IRCTC live train availability and Tatkal opening windows
niffler travel trains NDLS MMCT
# 4. Run ITR Tax Optimization with Form-16 breakdown
niffler itr 1250000 --tds 75000 --sec80c 150000 --sec80d 25000 --hra 120000
# 5. Inspect EPF passbook and 30-year compounding retirement projection
niffler epf 60000 --tenure 4.5 --balance 380000
GitHub Actions CI/CD Integration — Use in Any Project
Deploy automated flight fare alerts or travel budget watchers into any repository workflow using the reusable composite action:
⚡
Marketplace Action Identifier
Reference in your workflows via uses: shubhransh-gupta/niffler@v1.
Salaried individuals filing ITR-1 (Sahaj) only need their Form-16 values. Niffler computes tax liabilities under both the New and Old Tax Regimes, factors in standard deductions and 87A rebates, and generates the exact schema-compliant ITR-1 JSON file. You can simply upload this JSON directly to incometax.gov.in with zero CA service charges.
Niffler adheres strictly to IRCTC regulations and Section 143 of the Railways Act. It does not bypass CAPTCHAs or use illegal bots. Instead, it serves as a compliant advisor: tracking live berths, displaying exact Tatkal opening times (10:00 AM for AC, 11:00 AM for Non-AC), and preparing passenger manifests for fast, official 1-click checkout.
Never. Niffler auto-fills only non-sensitive passenger and booking parameters (Name, Age, Berths, Contact). It then hands off the session by opening the merchant's authenticated payment gateway page directly in your browser, where you complete the payment yourself.